RY: It is hierarchical and handles disturbances (watch “Main video”).
RY: I think this would certainly be a possibility of something to aim at for our learning project.
BA: Impressive! However, looking a little deeper, I discovered that the project is all about developing physics-based computer animations rather than attempting to capture how real humans or other animals actually control their movements. PCT, in contrast, is all about understanding how the real systems work, based on their actual neuro-muscular systems. Given the âDeepLocoâ? projectâs goals, it is not surprising to find them using simulated âmusclesâ? that employ âoptimizedâ? insertion points (although they do show that models based on non-optimized insertion points will also work, although less well). The video showing behavior at various stages of learning reveals that good performance is not achieved until something on the order of 900 learning trials have been performed, so the learning algorithm does not seem to be particularly efficient, although it must be admitted that human learning of motor skills also requires a lot of practice. Algorithms that make use of complex computational techniques such as those employed in the DeepLoco project are unlikely to be implemented in our biological wetware. It seems that evolution has discovered powerful heuristic methods that achieve these goals without the mathematical complexities.
BA: That said, these researchers are in fact employing hierarchical control systems to carry out the simulated activities.
Yes, I did see it, Warren, and replied to it on CSGnet; however, I never received a copy of my post, so perhaps it got lost in cyberspace. Did anyone on CSGnet see it? If not, I’ll repost.
By the way, I’m getting two copies of this thread, one through the CC: and one through CSGnet. Perhaps the CC: list could be culled to include only those not subscribed to CSGnet? Alternatively, just send those who expressed an interest and not CSGnet as a whole.
Yes, I did see it, Warren, and
replied to it on CSGnet; however, I never received a copy of
my post, so perhaps it got lost in cyberspace. Did anyone
on CSGnet see it? If not, I’ll repost.
RY: It is hierarchical and handles disturbances (watch “Main video”).
RY: I think this would certainly be a possibility of something to aim at for our learning project.
BA: Impressive! However, looking a little deeper, I discovered that the project is all about developing physics-based computer animations rather than attempting to capture how real humans or other animals actually control their movements. PCT, in contrast, is all about understanding how the real systems work, based on their actual neuro-muscular systems. Given the âDeepLocoâ? projectâs goals, it is not surprising to find them using simulated âmusclesâ? that employ âoptimizedâ? insertion points (although they do show that models based on non-optimized insertion points will also work, although less well). The video showing behavior at various stages of learning reveals that good performance is not achieved until something on the order of 900 learning trials have been performed, so the learning algorithm does not seem to be particularly efficient, although it must be admitted that human learning of motor skills also requires a lot of practice. Algorithms that make use of complex computational techniques such as those employed in the DeepLoco project are unlikely to be implemented in our biological wetware. It seems that evolution has discovered powerful heuristic methods that achieve these goals without the mathematical complexities.
BA: That said, these researchers are in fact employing hierarchical control systems to carry out the simulated activities.
Yes, I did see it, Warren, and replied to it on CSGnet; however, I never received a copy of my post, so perhaps it got lost in cyberspace. Did anyone on CSGnet see it? If not, I’ll repost.
By the way, I’m getting two copies of this thread, one through the CC: and one through CSGnet. Perhaps the CC: list could be culled to include only those not subscribed to CSGnet? Alternatively, just send those who expressed an interest and not CSGnet as a whole.